Uncertainty-Aware Feature Mixing and Cross-Decoupled Pseudo Supervision for Semi-Supervised Semantic Segmentation

Abstract

Semi-supervised semantic segmentation aims to maximize the training performance for a limited annotation cost. Existing methods such as cross pseudo supervision have shown excellent performance, yet ignore potential information interactions between labeled and unlabeled data, and suffer from misleading incorrect pseudo labels. This paper takes two ways to improve each of these shortcomings. Firstly, we perform feature-level mixing and cross-decoupling using labeled and unlabeled data to establish potential interactions between the two types of data. Secondly, an uncertainty-aware loss re-weighting method based on information entropy is used to mitigate the negative effects of incorrect pseudo labels. Experimentally, our method further improves the previous cross pseudo supervision method with competitive performance on PASCAL VOC 2012 dataset under various data partition protocols.